Hacker News Reader

Comment by DeathArrow | Hacker News Reader

Is there a well known strategy game where AI doesn't have better chances to win than a human? Is there possibly, math wise, that such games exist?

I kind of dislike not having a chance when playing against bots.

View on HN

Don't play against bots then. There's plenty of people online to play against at any moment of the day.

Maybe I read the wrong forums but this particular hangup ("machines can do it better so what is even the point?") seems to be quite unique to chess as a hobby. I've never seen someone on an athletics tournament complaining that the sport is no longer fun because a cannon could throw the spear much farther than even the best athletes, or that they no longer enjoy going to the gym because a construction crane can lift much more than they will ever be able to. Chess is at its core a game of rigidly applying rules, pattern matching, and applying a depth-first search on as deep a search tree as you can manage. It should hardly be a surprise that computers can do that extremely well.

I appreciate your comment because I have thought a lot about chess cheaters who make all of their moves at the recommendation of a chess engine. When playing against them, you really don’t have a chance to win unless you’re playing on a fast time control. Online, and sometimes even in-person, it may not be obvious that someone is cheating which makes you feel all the more powerless. As an improving player though, it’s not a good thing to fixate on. But I can totally understand why some people would want to move on to some other game like maybe bowling, tennis, etc.

I find chess engines and any model that can play games themselves very fascinating though. I wonder why chess engines do not really play in a human-like way. That’s why I don’t even like to play against beginner bots because they will tend to make weird and obvious mistakes.

Can I suggest you consider why you’re playing?

We are taught from a young age that winning is the objective of playing a game, but what if it wasn’t?

What if it was to learn some hidden truth, or to explore your own tendencies to over-react, to become too rash, to become overly-defensive or to be disappointed when things didn’t always go your way? What if you took each game as a learning opportunity?

If you take a different perspective, bots and AI can help you a great deal - they aren’t going to feel smug about beating you, or remind you of it over a family dinner, or brag about it on social media. They can help you improve your gameplay and yourself.

I’m reminded of this when I think of recent AI advances in Go. The moves and strategies that emerged have left that community in awe, because they were open to learning from what AI produced, not just obsessed with trying to win.

That does not quite seem to be the point. There is a huge difference between "better" and "crushing".

> We are taught from a young age that winning is the objective of playing a game, but what if it wasn’t?

No, we aren't. The are multiple players in a game. So you win some, they win some and sure, you are taught to try to win more than 50%. But if you win 90+% of the time, you are wasting both players' time.

A handful of games have been invented with the purpose of giving humans an edge over AI. They may have had a chance when the best AI was alpha-beta search, but against modern reinforcement methods, I don't believe they have a chance.

The games invented were Octi [1] and Arimaa [2].

[1]. https://boardgamegeek.com/boardgame/450/octi

[2]. https://boardgamegeek.com/boardgame/4616/arimaa

A casual search shows some discussion of AI-resistant board games here:

https://www.reddit.com/r/boardgames/comments/eaept6/tell_me_...

And of course the famous Calvinball.

> "I kind of dislike not having a chance when playing against bots."

Conjures an image of the astronaut with a gun meme...

  - Wait, playing games to win is a waste of time now?

  - It always was.

There's a difference between playing to win and being frustrated because your opponent can consistently steamroll you effortlessly without even giving you a chance to set your game up.

Is there some shame in asking the AI to play worse?

It’s a machine, it won’t feel smug about knowing it could beat me if it wanted to.

I’ve never really understood people who are competitive about board games and similar inside-the-box scenarios. (There’s a certain “Dwight Schrute” aspect to that mentality that’s hard to watch.) I like losing in games because there’s no real-world cost to trying out some idiot strategy.

Civilization V? Warcraft 3?

But any answer will be such only because no one tried hard enough to make a supreme AI.

OpenAI's dota2 bots were pretty impressive, they even had an open public lobby for people to play against.

They didn't dominate against human pros, neither did alphastar. So Humans still have a chance there.

Play a high variance game of chance like cards or Magic the Gathering. Then when you win it's because of your skill and when you lose its because of chance. :)

I've found playing against weaker bots you lose on the sort of human nature of weaker players. In particular humans tend to have a bit of an internal narrative, but weaker chess bots tend to just throw in shitty moves.

It's pretty well known that bots resort to cheating to win at stuff like Civ. I'm sure there's a patch out there that removes the cheating, and you can stomp on computers there.

Civ evo [0] applies the same rules to AI and human players. It is supposed to have a very strong AI. Also, AI does not have a personality (friendly, aggressive, etc.); it just wants to win. Unfortunately, I could never bring myself to learn this new variant.

[0] http://c-evo.org/

The problem with video game bots in games like Civ is that "playing well" is not actually the core goal of the devs. Rather, it has to be enjoyable to play against and easy enough to build and debug in the limited time available until release. Given those constraints, most game devs just give the bot free resources and call it a day.

If somebody would bother to put in the time to make it, a well-crafted Civ bot could probably roflstomp every human out there just by virtue of perfectly micromanaging its production in every city on every turn without effort.

Yes. Risk. There's a large online community around it now and lots of YouTube videos of people playing it to a high strategic level. There's all sorts of signalling going on between players and bots are easy to counter even at the highest settings.

I would say any game that includes a lot of uncertainty, subjectivity and random chance would mean that an AI would not usually have an edge over a human. Maybe games like diplomacy, magic the gathering, pandemic, Gloomhaven, maybe poker (?), backgammon (?) etc.

On a sidenote, I think it’s interesting that we’re a point where it’s starting to get hard to come up with games that humans can beat AI at.

I would like to hear your reasoning as to why you think that a human is especially good at dealing with uncertainty. My experience is the opposite.

I’m not saying a human would be better than an AI. Just saying it would be a more equal playing field, since neither would have a significant advantage.

I think this depends on whether the human has a chance in a particular run (yes) or on average of many runs (probably no because the AI will calculate the probabilities better).

But in the extreme case of a random game (like rolling the highest number on a die) they are equal (obviously).

True randomness doesn't drive better results for humans, due to some intrinsic humane qualities, but rather eliminates the advantage of large volume data access and processing of AI. Basically true randomness levels the playing field. Backgammon is a perfect example.

You want a game where searching large amounts of data, computing moves and calculating probabilities fast doesn't help.

Maybe some randomness will help, but might not be enough.

My ideea is that bots can't win in the real world economy, no fund driven solely by algorithms can win more than funds driven by humans.

So if we can find a game modeled like the economy, where nothing is random but many things are uncertain, then it might be harder for the software to win.

I would imagine in many games randomness (e.g. through throwing dice, or pulling cards from a randomized pack) add noise, but the underlying strategy (including processing data, calculating probabilities etc.) still help. In such a game even though an AI might be superior, the randomness might mean that occasionally the human wins. Or for an extreme example, if nothing you do in the game really matters, it's all just down to random chance, then the AI:human win rate should approach 50%. But such a game would probably not be particularly enjoying to play.

But yes, I think you're right that you'd need a game where crunching a lot of numbers really fast doesn't give you an advantage. For instance, if the state space of the game is so big that number crunching is useless, and other approaches like AI style pattern matching (used IIRC by Alpha-Go?) don't work either.

Though ultimately, what is the uniquely human trait that would allow a human to beat an AI? Can you make a game that depends on that? Is there even such a thing?

> if nothing you do in the game really matters, it's all just down to random chance... But such a game would probably not be particularly enjoying to play.

I guess all the slot machine players tend to differ in opinion there :)

Sure, slot machine operators adjust winning chances so players keep on playing, but to a player, it's not influenced by anything they do, other than "just one more time and I'll win".

TBH, "real world economy" does allow humans to rewrite the rules, and they've done so: issuing more bonds, deflating the currency, printing more money, bailing out broke banks, hiding facts and selling before downturn goes public, pure and simple fraud...

>I would like to hear your reasoning as to why you think that a human is especially good at dealing with uncertainty.

We can ask any successful CEO. Fortune 500 companies would use bots if that was possible.

Uncertainty doesn't equal randomness. Randomness is flipping a coin and asking you the result. Uncertainty is hiding the coin behind my back and asking you in which hand is it.

That's disanalogous to board games. We're comparing board games with uncertainty to board games without uncertainty. In either of these categories, the thing that makes AI competent is unlimited training data due to self-play.

Scaringly enough, Diplomacy seems to work well for AI too. https://ai.meta.com/blog/cicero-ai-negotiates-persuades-and-...

Computers are better than humans at poker (DeepStack) and backgammon (eXtreme Gammon). XG for example is commonly used by expert backgammon players to analyse play, much like how engines are used in chess.

There is no reason why computers wouldn't eventually beat a human in the others, if someone writes a narrow AI for them. Consider for example, AlphaStar for StarCraft.

>I would say any game that includes a lot of uncertainty, subjectivity and random chance

I would thought the same haven't I been reading articles on how bots won against humans at poker.

I am not sure about something like backgammon or monopoly.

There will never be anymore. AI simply outperforms humans at increasing rates. When in doubt, for a specific game a narrow AI can be trained up.

For the current 1-2 years you can have a shot at physical/sports games, since robots are lagging behind AI capabilities, quite a lot.

Robots can’t even carry enough energy to play a physical/sports game, even if their AI was perfect and cost zero energy itself. There’s no imminent tech that will change this. We have many years.

I’d love to be wrong about that, but I’m probably not.

Check out Robocup for superhuman speed miniature soccer games. But the games are small and short. The humanoid games are much less well developed.

Probably we differ here on the timespans, thats okay :-)

Personally I explicitly do not expect robots to actually have human-like shape but rather be optimized for the thing they do, which can change the energy requirements or otherwise things that might look "unfair" to humans. Usually this is one of the key reasons where different timelines come from in discussions, where people assume that a humanoid robot needs to be build that does the same movements like humans but somehow better/faster/... while I differ on that point.

That's moving the goalposts (pun intended :)): having a wheeled robot with an enclosed (ball protection!) basketball holder slide around the basketball court and extend the basketball holder up to the rim would certainly allow it to beat humans, but it wouldn't really be "playing basketball".

We engage in sports because they are fun and explore the limits of our physical and mental abilities (or well, go above our limits, as all the kinesio-tapes and supplements in pro sports indicate). Involving robots can only continue to be fun if robots work with roughly the same restrictions.

If we want to measure AIs against our mental abilities, it's only fair to pit them with controlling as much "machinery" and suboptimal movements as humans do in a particular sport: that versatility is what makes our intelligence so amazing and, well, "general".

but thats the same as in Chess, Go or even StarCraft: the AI is absolutly impossible to beat for human players unless it has explicit flaws built-in.

The general point I want to make is that it doesn't make sense trying to compare humans with AI (and somewhat soon) robots, because the will outperform everyone. Handycapping them is not a solution either. The "amazing versatility" of humans also will not persist for too long anymore I'd say, therefore the only way to have a real and fair comparison/pit is against other humans in the future.

Well, let's agree to disagree about versatility: I don't see non-specialized AI outperforming humans in a number of years (a human will easily beat chess AI at Go).

While I am somewhat impressed by the conversational acumen of recent LLMs, the fact they can produce outright garbage tells me we are as far away from synthesising multiple types of intelligences as ever: humans simply need far smaller input data set to start recognizing patterns and rules (as witnessed by kids learning to speak).

I am not yet convinced we'll live to see something like actual self-driving cars with as much capability as an average human driver: I believe augmented environments are needed (communicating roads, signals and cars) to get to self-driving cars.

The one thing they've got going for them is consistent focus, whereas practical ability in humans significantly depends on the mood, tiredness, level of multitasking...

> the will outperform everyone

I'll try to make the same point land a bit harder:

Robots will not compete any time 'somewhat soon' in many/most physical tasks, because their energy storage capability, motor efficiency, strength to weight ratio, and many other mechanical and sensing properties are not good at all compared with humans and other animals, and can't be fixed by software. Everyday intuition underestimates how large this gap is.

Brains are only part of the requirement for competence in the real world.

I feel like it is possible, but just not anything we will pursue. Nuclear reactors? Inductive charging in the floor? some kind of overhead electric like we use to power trains? I'm sure it's "possible", but probably not safe for humans to be around.

World of Warships has lots of strategy to it (and the bots suck at it). Good players don't shoot where the ship would be if it kept going straight but take into account the map, other ships positions and psychology to guess evasive maneuvers.

> and the bots suck at it

It is important to distinguish two thing: "Computers can't solve this problem" vs "the currently existing bots suck". One is when the best programers trying their darndest can't make a bot good, while the other is when nobody even really tried.

(For example I can reliably run faster than Usain Bolt when he is not made aware that we are racing. :P )

This case sounds very much the second kind. There can be many reason for this, the simplest is that the makers of the game don't win more players by absolutely crushing them. It is a case where after a certain level of advancement you don't earn more money by making better bots. In fact you might start earning less if you take away the fun of the game. And of course anti-cheating measures will hamper 3rd party bots if they are not explicitly invited, and why would they be?

Simply saying estimating which way someone will evade / evading randomly doesn't sound like the kind of thing a computer would be bad nowadays. If someone tries that is.

come join us at Beyond All Reason :)

https://www.beyondallreason.info/

or if you're an AI maker come join us also and make a better bit for us to play against :)

cough https://zero-k.info/ cough

Most board games don't have any superhuman AI. You could play race for the galaxy for example. There are bots, but you can achieve very high win rates against them.

You can adjust bot difficulty. If that doesn't work for you nothing will because with humans you also can never ever beat the best player in the world.

starcraft broodwar is one such game, although the AI bot has to be handicapped to have the same visual and APM constraints as a human player.

If the bot gets to have APM in the 400-500 range and has full map vision, even the top pros will get smoked no contest.

Reply on news.ycombinator.com